KROVEX: Multimodal Graph Fusion with Statistically Guided Parsimonious Descriptor Selection for Molecular Property Prediction

Environment installiation

• This code was tested with Pytorch 2.1.0, cuda 12.1, torchvision 0.16.0

• Download ananconda/miniconda if needed

• Create an environment:

  conda create -n krovex python=3.8

• Activate conda

  conda activate krovex

• Install Pytorch:

  pip install torch==2.1.0+cu121 torchvision==0.16.0+cu121 torchaudio==2.1.0+cu121 -f https://download.pytorch.org/whl/torch_stable.html

• Install dgl library:

  pip install dgl==2.2.1 -f https://data.dgl.ai/wheels/repo.html

• Install packages using the requirement file:

  pip install -r requirements.txt

Run the main code

• To run the main code: python main.py

• To run the code on only a few batches, epochs, and folds, you can change them in: .\configs\config.yaml

Use KROVEX on a new dataset

To implement a new dataset, you need to select descriptors through a Descriptor Selection. Check \descriptor_selection folder.

Descriptor Selection

• To run the code main_descriptor_selection.py, you may need some preparation:

  • Install rpy2 libarary: conda install -c conda-forge rpy2

  • Download R

  • Check your version of R: r --version. The results should be the same version of R you downloaded.

  • Specifiy R path in .\configs\config.yaml

• Run the code for descriptor selection:

  python .\descriptor_selection\main_descriptor_selection.py

• To incoporate descriptors into the model, check utils\mol_collate.py and utils\mol_conv.py for details.